Volume 24, N°2
- proof for author
- 0.1
J-ottings 52
All but one
by Norman Thomson
A frequent requirement in applied mathematics and statistics is to evaluate sums and products omitting just one variable or dimension. The notion of ‘all but one’ can be interpreted in two ways, depending whether the ‘one’ is to be systematically omitted, or obtained by a merge with an existing dimension, that is by reduction.
Retaining all but one
As an example of the first case, adding or multiplying ‘all but one’ items
in a list progressively can be done by using the hook
invf~f/ which means (f/x)invf x,
where invf is the inverse verb of f,
for example:
i.5 0 1 2 3 4 (-~+/)i.5 NB. sum five items omitting one at a time 10 9 8 7 6 (%~*/)1+i.5 NB. multiply five items omitting one at a time 120 60 40 30 24
This extends readily to objects of higher rank:
]b=.i.3 4
0 1 2 3
4 5 6 7
8 9 10 11
(-"1~+/)b NB. sums of all rows but one
12 14 16 18
8 10 12 14
4 6 8 10
(-"2~+/"1)b NB. sums of all columns but one
6 5 4 3
18 17 16 15
30 29 28 27
The rule is that the rank operand for the inverse verb should be one greater than that of the inserted verb.
Generalising ‘retaining all but one’
Another tool which deals with ‘retaining all but one’ is the suffix adverb which eliminates n items from a list in all possible ways:
(1+\.i.5);(2+\.i.5);(3+\.i.5) +-------+-----+---+ |1 2 3 4|2 3 4|3 4| |0 2 3 4|0 3 4|0 4| |0 1 3 4|0 1 4|0 1| |0 1 2 4|0 1 2| | |0 1 2 3| | | +-------+-----+---+
In the J line above + is monadic, which for real numbers is
a ‘do nothing’ verb; that is, the left arguments 1, 2 and 3 are not
added to elements of i.5 but are arguments of the derived
verb +\. that indicate how many items are to be dropped,
progressively working from the left. The first case with 1 as left
argument is thus the ‘all but one’ case.
An application of this is the calculation of minors of a determinant.
Consider the rank-2 object z:
z 3 6 7 9 3 2 9 7 7 (1&(+\.))"2 z NB. select 'all but' one rows 9 3 2 9 7 7 3 6 7 9 7 7 3 6 7 9 3 2
Now combine the structural verb transpose with the adverb suffix to switch rows and columns for all but one row at a time:
transuff=.1&(|:\.)"2 <"2 transuff z +---+---+---+ |9 9|3 9|3 9| |3 7|6 7|6 3| |2 7|7 7|7 2| +---+---+---+
and the result is a list of the transposed planes of
(1&(+\.))"2 z.
Do this twice using the power adverb to generate three boxes within each
of the above boxes; the row/column switch is fortuitously reversed and
the minors of z obtained:
<"2 transuff^:2 z +---+---+---+ |3 2|9 2|9 3| |7 7|9 7|9 7| +---+---+---+ |6 7|3 7|3 6| |7 7|9 7|9 7| +---+---+---+ |6 7|3 7|3 6| |3 2|9 2|9 3| +---+---+---+
Define
minors=.transuff^:2 NB. minors unboxed det=.-/ .* NB. determinant
The determinants of the minors are given by
each=.&> ]dz=.det each <"2 minors z 7 45 36 _7 _42 _33 _9 _57 _45
This is verified by using the verb det dyadically:
(det z);dz det |:z
+-+------+
|3|3 0 0|
| |0 _3 0|
| |0 0 3|
+-+------+
It is often convenient to use cofactors, that is the signed determinants of minors. This requires multiplication by a matching matrix whose diagonals are alternately +1 and –1. One way of obtaining this matrix is:
signs=.monad : '-_1+2*2|+/~i.#y.'
so that
cof=.signs * det each @<"2@minors cof z 7 _45 36 7 _42 33 _9 57 _45
To summarise this section:
transuff=.1&(|:\.)"2 NB. transpose with suffix minors=.transuff^:2 NB. minors unboxed det=.-/ .* NB. determinant each=.&> signs=.monad :'-_1+2*2|+/~i.#y.' NB. alternate 1,_1 cof=.signs * det each@<"2@minors NB. cofactors
Reducing all but one
Rank again is at the heart of the matter, especially as typical
experimental data is structured so that dimensions correspond to
variables. However, J inherently structures higher-rank data as lists,
then lists of lists, lists of lists of lists and so on, which implies a
nesting within dimensions which is not usually reflected in the real
world variables to which the dimensions correspond. For definiteness
use q to demonstrate:
]q=.i.2 3 4 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
The standard definition of rank is:
rk=.#@$ NB. rank
+/ ‘inserts’ + between items of the list at
the topmost level, thereby reducing rank by one, that is merging planes:
+/q NB. equivalent to +/"n q for n>2 12 14 16 18 20 22 24 26 28 30 32 34
Explicit rank conjunctions allows such reduction to take place at any level in the rank hierarchy:
(+/"1 q);(+/"2 q) NB. merge cols ; merge rows +--------+-----------+ | 6 22 38|12 15 18 21| |54 70 86|48 51 54 57| +--------+-----------+
Progressive reduction to the level of a rank-1 object is obtained using a recursive verb:
sum=.sum@(+/)` ] @.(=&1@rk) NB. sum down to rank 1 sum q 60 66 72 78
The above values are readily verifiable as the column sums of
q. To find other such sums, say row sums, transpose the data
so as to bring the corresponding dimension to the top level. This suggests
a general verb which takes the list i.n and performs the a
set of n possible shifts necessary to bring each item in turn into the
leading position:
EACH=.&.> shifts=.|.EACH< NB. all distinct shifts of a list shifts i.rk q +-----+-----+-----+ |0 1 2|1 2 0|2 0 1| +-----+-----+-----+
If the argument to shifts is a list generated by i., the
result is left arguments to transpose which provide all the restructured
forms of q to which sum can be applied. This in
turn is determined by the rank of the data matrix, so define
targs=.shifts@(i.@rk) NB. arguments for |: targs q +-----+-----+-----+ |0 1 2|1 2 0|2 0 1| +-----+-----+-----+
The full set of transpositions to supply all possible sums by dimension is then
transposes=.targs |:EACH < NB. reqd. transposes transposes q +-----------+-----+--------+ | 0 1 2 3| 0 12| 0 4 8| | 4 5 6 7| 1 13|12 16 20| | 8 9 10 11| 2 14| | | | 3 15| 1 5 9| |12 13 14 15| |13 17 21| |16 17 18 19| 4 16| | |20 21 22 23| 5 17| 2 6 10| | | 6 18|14 18 22| | | 7 19| | | | | 3 7 11| | | 8 20|15 19 23| | | 9 21| | | |10 22| | | |11 23| | +-----------+-----+--------+
These values are readily verifiable by summing the columns in the boxes above:
sum EACH@transposes q +-----------+------+---------+ |60 66 72 78|66 210|60 92 124| +-----------+------+---------+
To give these sums in dimension order, that is so that the
$EACH of the result matches the shape of q,
write
allsums=.1&|.@(sum EACH@transposes) allsums q +------+---------+-----------+ |66 210|60 92 124|60 66 72 78| +------+---------+-----------+
To summarise this section:
rk=.#@$ NB. rank sum=.sum@(+/)`] @.(=&1@rk) NB. sum down to rank 1 shifts=.|.EACH < NB. all shifts of i.n targs=.shifts@(i.@rk) NB. arguments for |: transposes=.targs |:EACH < NB. reqd. transpositions allsums=.1&|.@(sum EACH@transposes)
For comparison, here is a one-liner picked from the J Software forum some
years ago which performs the same function by using the power adverb
^: to apply +/ the requisite number of times
with transpositions required between steps as in the version above:
mt=.(<@(+/^:(<:@rk)@:|:)"0 _~i.@rk) mt q +------+---------+-----------+ |66 210|60 92 124|60 66 72 78| +------+---------+-----------+
Subtotaling
It can be useful to be able to append such reductions to the original data as in:
total=.,+/ NB. append totals for leading dimension sub=.3 :0 i=.0 [ r=.total y. while.(i<<:rk y.)do.r=.total"i r [ i=.i+1 end. ) sub q 0 1 2 3 6 4 5 6 7 22 8 9 10 11 38 12 15 18 21 66 12 13 14 15 54 16 17 18 19 70 20 21 22 23 86 48 51 54 57 210 12 14 16 18 60 20 22 24 26 92 28 30 32 34 124 60 66 72 78 276
Multi-statement lines using [ as a separator as in
sub allow something approaching the succinctness of tacit
definition. However the individual statements are executed from the right
since [ itself is just another verb. It is easy to remember
that it is [ rather than ] which is the
separator, since, for example, it is 0 to the left which is assigned to
i in the first line of sub. As far as the J
interpreter is concerned it is really only one line which is executed;
multiple lines are essentially just an orthographic device.
